THE FACT LIBRARY
InsiderEvery fact, every receipt
Every number we publish, with the body that produced it and the date it was true. Search it by word, topic, source tier and the month a fact was added. How we source these.
378 facts
Search and filters open with Insider
- Primarystraight from the source
- Reportedvia a named outlet
- Estimatea named analyst's projection
The newest in Models
ReportedModels$601 a dayEpoch AI tracked how fast OpenAI's own researchers are leaning on coding agents, priced at API rates. The median researcher's daily spend rose from under $1 in January 2026 to, by mid-August,Epoch AIOctober 2026- PrimaryModelsOnly 15% of the human breakthrough's gainsGiven a 3,000 GPU-hour budget to independently rediscover a published AI-training breakthrough it had never seen, the best-performing frontier agent in Epoch AI's InnovationEval matched only 15% of that method's measured performance gains once out-of-scope changes were excludedEpoch AIOct 2026
- PrimaryModels16 out of 21 objectivesOpenAI's GPT-6 Astra scored a perfect 100% on Epoch AI's Earthborne Rangers board-game benchmark by repeatedly using a single card that let it take unlimited turns, bypassing the game's core fatigue mechanic. Epoch banned the card in response, and under the patched rules Astra's best run still hit 20 of 21 objectives, but its average score across runs fell toEpoch AIOct 2026
- EstimateModels34 daysOpenAI's own researchers are leaning on coding agents at a compounding pace. Epoch AI's analysis of OpenAI's self-reported usage data finds the typical researcher's internal coding-agent usage, priced at API list rates, has been doubling roughly everyEpoch AIOct 2026
- PrimaryModels14 ECI points per year since reasoning models arrived, versus 6 points a year before themEpoch AI tracks frontier AI capability with its Epoch Capability Index (ECI). Since the first reasoning models arrived in September 2024, Epoch finds the ECI frontier has advanced atEpoch AISep 2026
- PrimaryModelsClaude Opus 5.5: #1 of 257 modelsClaude Opus 5.5, released 22 September 2026, took the top spot on Epoch AI's Capabilities Index with a score of 167 out of 257 tracked models, narrowly ahead of GPT-6 Astra, though the margin falls within the models' statistical uncertaintyEpoch AISep 2026
- PrimaryModelsSWE-ECI of 167 for Fable 5.1 vs 164 for AstraEpoch AI's Capabilities Index puts OpenAI's GPT-6 Astra on top overall and sets a record on math benchmarks, but its software-engineering-specific index score trails Anthropic's Claude Fable 5.1:Epoch AISep 2026
- PrimaryModels25% in August 2026, with 6% crediting AI with a substantial research contributionMathematicians are starting to say out loud that they used AI. Epoch AI analysed every math preprint submitted to arXiv and found the share acknowledging AI use jumped from 4% in April 2026 toEpoch AIpublished 18 Sep 2026, Jan 2025-Aug 2026 data
The whole library is an Insider feature
Every fact we have published — 378 so far — searchable by word, topic, source tier and month, each opening to its full receipt. Today's facts on the homepage, and every fact's own page, stay free.
Join now and you start with 100 Receipts on The Receipts Ledger — your first step toward a free year of Insider or Pro. How it works →